Node: crypto Decipher: setAuthTag() must be called before update()

[achronos0]

• Version: v8.11.3
• Platform: Darwin foo.local 17.7.0 Darwin Kernel Version 17.7.0: Thu Jun 21 22:53:14 PDT 2018; root:xnu-4570.71.2~1/RELEASE_X86_64 x86_64
• Subsystem: crypto

crypto Decipher docs say that setAuthTag() must be called before final().
In fact setAuthTag must be called before any call to update(), in GCM mode at least.

If you call update() before setAuthTag(), final() will throw regardless of the correctness of the ciphertext and tag.

Test case:

const authTagBeforeUpdate = false; // true works, false throws

const crypto = require('crypto');
const cleartext = Buffer.from("abcdefghijklmnopqrstuvwxyz0123");
const key = Buffer.from("abcdefghijklmnopqrstuvwxyz012345");
const iv = Buffer.from("0123456789ab");
console.log('cleartext', cleartext);
console.log('key      ', key);
console.log('iv       ', iv);

const cipher = crypto.createCipheriv('aes-256-gcm', key, iv);
const ciphertext = Buffer.concat([cipher.update(cleartext), cipher.final()]);
const authTag = cipher.getAuthTag();
console.log('authTag  ', authTag);

const decipher = crypto.createDecipheriv('aes-256-gcm', key, iv);
if (authTagBeforeUpdate) {
    decipher.setAuthTag(authTag);
}
let result_update = decipher.update(ciphertext);
if (!authTagBeforeUpdate) {
    decipher.setAuthTag(authTag);
}
let result_final = decipher.final();
let result_cleartext = Buffer.concat([result_update, result_final]);
console.log('result   ', result_cleartext);

I suggest changing the API docs to make it clear that the auth tag must be supplied before any data is read in. setAuthTag() must be called before final() and should be called before any call to update().

Second: is this actually the desired behaviour?

As far as I can tell, there is no requirement for GCM to know the auth tag prior to deciphering data. OpenSSL's own example provides it after EVP_DecryptUpdate, before EVP_DecryptFinal_ex (https://wiki.openssl.org/index.php/EVP_Authenticated_Encryption_and_Decryption#Authenticated_Decryption_using_GCM_mode).

In cases where the auth tag is appended to the ciphertext (e.g. Java), as things stand the entire ciphertext must be buffered. For large data this is a serious problem. It would make more sense to decrypt chunks as they are received, treating them carefully and discarding everything if the auth tag eventually proves to be incorrect.

[achronos0]

OK reading the source I can see how and why this behaviour occurs, and I have found a workaround, maybe docs could mention it.

From master/src/node_crypto.cc

• function CipherBase::SetAuthTag does not actually set anything in the engine (it does not call EVP_CIPHER_CTX_ctrl)
• function CipherBase::Update is where the tag is actually set, the first time it is called after the key is set (it calls EVP_CIPHER_CTX_ctrl(... EVP_CTRL_AEAD_SET_TAG ...))

So the result is that:

• a call to (js) setAuthTag() after all calls to update() and before calling final() doesn't work, the tag is never actually set in OpenSSL
• multiple calls to setAuthTag() e.g. setAuthTag() then update() then setAuthTag() then final() doesn't work, only the first tag is actually set in OpenSSL, the second tag is ignored

The workaround I found is to set the authtag then make one final call to update() with empty data. That works.

My example above, modified with workaround so it works in both cases:

const authTagBeforeUpdate = false; // true works, false now also works

const crypto = require('crypto');
const cleartext = Buffer.from("abcdefghijklmnopqrstuvwxyz0123");
const key = Buffer.from("abcdefghijklmnopqrstuvwxyz012345");
const iv = Buffer.from("0123456789ab");
console.log('cleartext', cleartext);
console.log('key      ', key);
console.log('iv       ', iv);

const cipher = crypto.createCipheriv('aes-256-gcm', key, iv);
const ciphertext = Buffer.concat([cipher.update(cleartext), cipher.final()]);
const authTag = cipher.getAuthTag();
console.log('authTag  ', authTag);

const decipher = crypto.createDecipheriv('aes-256-gcm', key, iv, { authTagLength: 16 });
if (authTagBeforeUpdate) {
    decipher.setAuthTag(authTag);
}
let result_update = decipher.update(ciphertext);
if (!authTagBeforeUpdate) {
    decipher.setAuthTag(authTag);

    // Workaround:
    decipher.update(Buffer.from([]));
}
let result_final = decipher.final();
let result_cleartext = Buffer.concat([result_update, result_final]);
console.log('result   ', result_cleartext);

[addaleax]

/cc @nodejs/crypto

[tniessen]

I just noticed that there is still an edge case where this doesn't work: If a user restricts the permitted length of GCM authentication tags to a single value while calling createDecipheriv via the authTagLength option and sets the authentication tag after calling update, verification will still fail. Working on it.

[tniessen]

Fixed in #22828.

[bencmbrook]

Is there any way to use a decipher stream, and set the authTag later? In other words, perform the integrity check after decipher is done.

It seems like one still needs to set the authTag _before_ calling .pipe(decipher), since the stream will automatically call decipher.final() when the stream's 'end' event fires.

[achronos0]

@bencmbrook that is my original use case in this ticket as well, and yes you can.

You will need to use your own custom decipher stream, so you can set authTag before the stream finishes.
But -- you probably already need to have a custom stream, yes? In order to parse out authTag from the encrypted data?

In my case I use a Transform stream as a wrapper around the crypto stream, so my transform pre-parses the input before passing it on to the crypto decipher stream, and when it knows its found the authTag and has read all of it, it then calls decipher.setAuthTag() and then decipher.final().

[bencmbrook]

Thanks @achronos0 - put some basic source code together for the next person :)
https://github.com/transcend-io/lazydecipheriv/